A clustering approach for translationese identification

Proceedings of Recent Advances in Natural Language Processing, pages 532–538, Hissar, Bulgaria, 7-13 September 2013. A clustering approach for translationese identification Sergiu Nisioi Faculty of Mathematics and Computer Science, University of Bucharest sergiu.nisioi@gmail.com Liviu P. Dinu Centre for Computational Linguistics, Bucharest ldinu@fmi.unibuc.ro Abstract Our paper is concerned with investigating the impact of translationese on the novels of a bilingual writer and asking whether one could determine the authorship of a translated document. The main part of our paper will be centered on selecting a good set of lexical features that can be con- sidered characteristic for an author. We used in our research the novels of Vladimir Nabokov, a bilingual author, who wrote his works in both Russian and English. Each text is represented by a vector of function words. We are interested in deter- mining how the results vary across differ- ent feature sets and which feature set could be considered the most representative. In order to inspect our results we used a hier- archical clustering method and draw con- clusions based on the most frequent result. 1 Introduction The term “translationese” proposed by Geller- stam (1986) currently means the entire sum of lin- guistic characteristics (Hansen, 2003) that a trans- lation exhibits in comparison to a text written na- tively in a language. The existence of transla- tionese has been discussed and more recently var- ious methods (Koppel et al., 2011; Ilisei et al., 2010) for identifying translationese have been de- vised. In the same context, an interesting discussion regards the equivalence in style between the trans- lated and original text. As Boase-Beier (2006) suggests, among other factors, the stylistics of a translation is highly related to the choices made by the translator in re-creating the original style, the translator having a specific “fingerprint” (Wang and Li, 2012). Our concern is investigating the impact of translationese on a bilingual writer and asking whether one could determine the author- ship of a translated document. The problem of au- thorship attribution is postulated on the grounds that the human stylome exists. The stylome is defined as “a linguistic fingerprint that can be measured, is largely unconscious, and is con- stant” (van Halteren et al., 2005). A fairly large amount of literature is dedicated to authorship problems and extensive overviews are provided by Juola (2006) or Stamatatos (2009). We are mostly interested in finding the lexi- cal features that can be used to discriminate or to characterize original and translated documents and once these words are presumably found, what is their role in authorship attribution for such docu- ments? The main part of our paper will be cen- tered on selecting a good set of lexical features to detect translations in a corpus of original doc- uments. In order to investigate our problems we have constructed two corpora from the novels of Vladimir Nabokov: a Russian corpus containing original Russian works and translations from En- glish and a second corpus containing English orig- inal works and translations from Russian. The details …

Content truncated to 500 words